[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-speed-up-3d-vision-ai-models-sixfold":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},11315,"researchers-speed-up-3d-vision-ai-models-sixfold","Researchers Speed Up 3D Vision AI Models Sixfold","GeoGR compresses redundant 3D scene tokens and skips irrelevant computation, speeding up inference while keeping most of the model's accuracy.","A new technique trims the fat off 3D vision-language models, running them over six times faster without gutting their accuracy.\n\nResearchers describe GeoGR, a two-stage framework for compressing how AI models process 3D scenes captured from multiple camera views. The first stage, GeoSemZip, merges repeated observations of the same scene regions across views into compact 3D voxel tokens, keeping the most informative ones and folding the rest into background anchor features so nothing is fully discarded. The second stage, GroupRoute, skips low-relevance tokens past entire layers of the language model, letting them rejoin the computation later at designated checkpoints. Trained on just 67,000 samples - 30% of the original dataset - GeoGR kept 95.2% and 96.8% of the original models' performance on two different 3D VLMs while speeding up prefill by 6.8x and 6.1x.\n\nThis matters because multi-view 3D scene understanding is one of the more expensive jobs you can hand a vision-language model - think robots, AR headsets, or indoor-mapping tools that need to reason about space, not just pixels. Most efficiency tricks for these models either throw away information or add complexity that eats into the savings; GeoGR's approach of routing tokens in and out of computation, rather than deleting them outright, is a more surgical way to cut cost.\n\nWhether a 6x prefill speedup survives contact with real deployment hardware and messier real-world scenes - rather than benchmark datasets - is the open question here.","[\"ai\",\"vision-language-models\",\"token-compression\",\"3d-scene-understanding\"]","2026-10-12T04:00:00.000Z","2026-10-12T07:36:55.895Z","2026-10-12T07:37:00.786Z","published",null,[],"ai",[24,26,27,28],"vision-language-models","token-compression","3d-scene-understanding",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.13035",0,{"sections":35},[36,39,44,49,54,59,64,69,74,79,84,89,94,99],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",6966,{"name":40,"slug":41,"count":42,"latest_published_at":43},"Security","security",952,"2026-10-12T04:54:12.000Z",{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",496,"2026-10-11T20:06:43.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",489,"2026-10-12T13:30:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",237,"2026-10-12T13:24:45.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Science","science",196,"2026-10-11T15:34:40.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Consumer Tech","consumer-tech",192,"2026-10-12T14:07:13.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Startups","startups",121,"2026-10-12T13:22:34.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":78},"Software","software",115,"2026-10-11T07:34:45.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Dev Tools","dev-tools",111,"2026-10-12T11:22:24.000Z",{"name":85,"slug":86,"count":87,"latest_published_at":88},"General","general",70,"2026-10-11T14:15:00.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"Gaming","gaming",59,"2026-10-09T11:43:43.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Reviews","reviews",35,"2026-10-12T14:00:00.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]